YOLO Instance Segmentation Model Comparison for Drone Detection as Visual Servo Control Marker

摘要

The development of collaborative aerial and ground robot technology offers a promising solution for volcano-exploring robotic systems. However, certain components are missing, particularly concerning the concept of ground robot retrieval. While the use of visual servo control is not new, it can be applied for aligning drones and rovers to facilitate ground robot retrieval. Traditional visual servo control methods often rely on fiducial markers, which may lack robustness in practical implementations. Within this paper, the authors propose an alternative visual servo control marker suitable for real-world scenarios. We present a comparative analysis of existing instance seg-mentation models from different YOLO versions, specifically, YOLOvS, YOLOv7, and YOLOv8, trained for drone detection. Additionally, we introduce a novel dataset, DroneCVS, comprising drone images captured from the ground robot's perspective under varying weather and altitude conditions. Our experimental findings demonstrate the instance segmentation model's capacity to acquire and learn drone segmentation from the DroneCVS dataset, achieving outstanding performance in different in evaluation metrics. In most evaluation scenarios, YOLOvS outperforms other models, leading to the conclusion that it provides the best balance between speed, weight, and performance in real-time applications.